0.20.3

Focus: one error behind different calls shows as a loop (first error line, numbers and paths taken out), and an error met in three or more sessions suggests a CLAUDE.md note; a turn with far more tool calls than this project's turns usually take is called out; the cache hint judges this turn's requests from the second on; re-reads reset at a compaction; the approval hint names this turn's calls only.

Bench: cost per session, each request priced from its tokens at its model's list price, so sessions read from transcripts get a figure too; tool calls per prompt.

Done gate: Fable is asked only when it can judge, a stop it did not judge is no pass, and after a failed answer it is left alone for five minutes; a declined call is no tool error; /agentmon clear mid-turn resets the whole turn; a turn under way at a plugin reload keeps its clock and gate.

Learning: a turn is saved even when the engine gives no usage; the stored copy is trimmed; /agentmon learn parses each transcript once, backfills error signatures, and writes only keys that survive the prune; a half-done migration no longer overwrites a newer key; git fetch names refs/heads/.